Gemini 4 Argon vs Muse Wonder: The Real Cost

0:00 · Two models, one point apart

So for about a fortnight, the two most talked-about AI models on the planet were basically tied. Google’s Gemini four Argon scored fifty-three on the Artificial Analysis Intelligence Index. Meta’s Muse Wonder scored fifty-two. One point. Which, honestly, is about the difference you get from running the same test twice. But one of them costs less than half as much to actually use. About one pound fifty a task, against three pounds forty. And that sounds like the end of the conversation. Cheaper model, same brains, job done. It isn’t. Because Google’s price has a sell-by date. Meta’s biggest move this month isn’t even the model you can buy. And the number that really decides your bill isn’t the one on the pricing page. I’ll get to all three. But first, let’s be clear about who just walked into this fight.

0:51 · Meta walks in

Fourteen months ago, Meta basically tore up its AI effort and started again. New lab, new stack, and a restructure led by Alexandr Wang, the founder of Scale AI. Since then, they’d shipped Muse Spark, a lightweight model built to be quick, and a Muse agent. Useful. But nothing that could sit at the top table. On the twelfth of October, that changed. Muse Wonder, and Muse Wonder Full. TechCrunch called it Meta’s first real attempt at a frontier model. And it was trained at Prometheus, Meta’s gigawatt-scale data centre in Ohio. Now, Meta’s pitch isn’t really, we’re cleverer than everyone. It’s two features with very Meta names. Active Memory, which is remembering what matters to you across conversations. And Active Focus, which is basically not giving up when a task gets hard. Trying another route, recovering from errors, and checking in before it does anything important. And Meta’s launch demos lean right into that. A playable kitchen game built from a few photos. A running plan worked out across your apps. Custom merch designed from your Instagram. And a wedding seating chart. Which, if you’ve ever had to do one, is genuinely the hardest problem in artificial intelligence. One of their research scientists put it like this. We put a tremendous amount of effort into making sure that our model can remember what’s important to you over days, weeks, months, and beyond. Fine. Every lab says something like that. So, receipts. Artificial Analysis, who are independent, scored Wonder Full at fifty-six. That’s eight points up on Muse Spark one point three, which came out just six weeks earlier. The biggest jump was on Terminal Bench, which tests command-line coding. Thirty-three percent, to fifty-seven. That’s not a tweak. On coding agents, Wonder Full scored sixty-eight, level with Claude Code running Sonnet five point five. Meta’s own testing also puts Wonder ahead of Claude, GPT six and Fable on its memory benchmark. Vendor numbers, so pinch of salt. But here’s the bit I didn’t expect. Knowledge is still a relative weakness for Wonder. Yet it pushed its score up mostly by making things up less. Its hallucination rate dropped from thirty-three percent to twenty-three.

2:58 · Google’s locked door

Now, Google. Argon was announced on the thirtieth of September. And then, well, you couldn’t use it. For two weeks it sat behind something called the Fairwind Program, a gated rollout for approved partners, plus a government pre-release review. So for a fortnight, the best-value model on the leaderboard was, as one router company put it, a set of numbers rather than an API. On the thirteenth of October, the door opened. The Gemini API, AI Studio, the Gemini app, and Pro and Ultra subscribers. Google’s pitching it at the heavy stuff. Complex coding, knowledge work, cybersecurity, and science. And the numbers are good. Fifty-three on the Intelligence Index, eighth out of two hundred and sixteen models. And first on the Vals Index, which weights finance, legal, tax and coding work by how much of the US economy they represent. Sixty-eight point nine percent. And this one I genuinely like. On a test of how often a model makes things up, Argon’s hallucination rate is fifteen percent. The figure for G P T six Astra is fifty-one. Artificial Analysis called it the lowest of any model in that top tier. Now, not everything’s a win. On Terminal Bench, the command-line test where Meta just jumped, Argon scores fifty-seven point four. Claude Opus five point five scores sixty-six point four. And that’s from Google’s own comparison table, so, fair play to them. Google’s internal claims are bigger again. More than a third more code submissions in their Antigravity tool, and help refactoring around six million lines of Chrome. Interesting. Unverified. File under nice if true. And, according to Business Insider, Google already has more Gemini four variants in testing, apparently called Barium and Carbon. Because of course they do. The periodic table is a long list.

4:43 · The hidden bill

Right. Money. This is where most of the coverage stops reading. Price per token first. Argon, two dollars per million tokens in, and ten dollars out. About one pound fifty, and seven pounds fifty. Muse Wonder, four in, and twenty out. Exactly double. The same as Claude Opus five point five. But nobody really pays per token. You pay per job. So Artificial Analysis ran both models through their entire index, and added up the bill. Argon, about one dollar ninety-nine per task. And for Muse Wonder, it’s four dollars fifty-two. That’s two point three times the cost, for one point less on the index. So why’s the gap so big? Here’s the bit that genuinely surprised me. Wonder is actually quite thrifty with its answers. About fifty-four thousand output tokens per task. Fewer than Argon’s sixty-two thousand. It’s the reading. Wonder pulls in about seven point seven million input tokens per task. That’s two to seven times what Argon and Open A I’s latest models use per task. Seventy-six percent of its bill is input. Now, most of that reading is cached, which means it’s billed at a big discount. Both models knock about ninety-five percent off cached input. But even at a discount, seven and a half million of anything adds up. So it writes less, but it reads a lot more. Which, if you’ve ever worked with someone who reads every single attachment before replying to an email, you’ll recognise. And now, the sell-by date. Argon’s price is introductory. It runs until the thirtieth of November. On the first of December, Google doubles it. Four dollars in, twenty out. Exactly Muse Wonder’s price. So run the same maths on cost per task, and Argon goes from about two dollars, to about four. Wonder is at four fifty-two. The gap that looked enormous, shrinks to about ten percent. Now, that’s my back-of-the-envelope, not a re-run of the benchmark. Caching changes things, and Argon was measured at high effort while Wonder was at max. But the direction is pretty clear. If you’re choosing on price, you’re choosing on a price that expires in six weeks.

6:51 · The free giant

And then Meta did something that, to me, is the bigger story. Three days after Wonder, on the fifteenth of October, they released the weights for Muse Spark Open. Eight hundred and sixteen billion parameters. Free to download. Free to fine-tune. Free to sell products built on it. With a couple of strings. You have to show, built with Muse Spark Open, somewhere in your product. And if you run it as a paid API business turning over more than twenty million dollars a year, you need a separate deal with Meta. Here’s the clever part. It’s what’s called a mixture of experts. There are two hundred and fifty-six specialists inside it, and for each word, only eight of them wake up. So of those eight hundred and sixteen billion parameters, only about forty-two billion are working at any one moment. Huge model, small active core. And people jumped on it. Within days, Hugging Face was already listing twenty quantised versions, and seven fine-tunes. Meta also says this open version had extra safety training, because once it’s out there, there’s no system around it to catch things. They recommend adding your own guardrails on top. Which is fair. Which is why Meta says it runs on two Mac Studios. Two. Not a data centre. Though, to be fair, they mean the M five Ultra ones with a frankly silly amount of memory. And the full-precision version still wants about two terabytes of graphics memory. So no, you’re not running it on the laptop you’re watching this on. But think about what that does to the whole price argument. Google is charging rent. Meta just handed out the keys to a very large house. And for a lot of businesses, the question stops being, which API is cheaper. It becomes, do we need an API at all.

8:30 · So which should you use?

So, if you’re actually choosing, and this is the sort of thing we test at AI Tools Review, here’s how I’d think about it. If you’re building something that reads a lot, long documents, big codebases, research, Argon, today, is the obvious pick. Lower hallucination, much lower cost per task. Just put a reminder in your calendar for the first of December. If you want a model that remembers you, and keeps going on long, messy tasks, that’s exactly what Meta built Wonder for. It also takes speech and video in directly, which Argon’s published inputs don’t. If you’re a coder, Wonder Full inside Muse Code is now level with Claude Code. But it’s subscription only. You can’t call Wonder Full from the API at all. And a tip if you do use Wonder through the API. It has an effort dial, from minimal right up to max, and every bit of thinking is billed as output. So run the simple stuff, extraction, sorting, tagging, at minimal, and save max for the jobs that genuinely need it. That alone can move your bill more than the choice of model. If you’ve got data you can’t send to anyone, or volumes that make API bills scary, Muse Spark Open is the interesting one. Not cheap to host. But nobody can double the price on you in December. And if the job is heavy terminal work, command-line agents, dev ops, the honest answer this week is neither. Claude Opus five point five still leads that test. One more thing. A one-point difference on an index is noise. Meta actually makes this point themselves, which I didn’t expect. Their blog talks about two models with similar scores, where one just handles unfamiliar tasks far better. They even use the internet’s words for it. Benchmaxxing, versus big model smell. When the AI labs start quoting Reddit at you, you know the leaderboard isn’t the whole story. Test on your own work. Run fifty real tasks through each one, and look at the bill, not the benchmark.

10:16 · The fight has moved

Zoom out, and the shape of this month is pretty clear. The top of the leaderboard is a crowd now. Google, Meta, OpenAI, Anthropic, all within a few points of each other. Both of these models read a million tokens at a time. Both cache. Both code well. On raw ability, the differences are getting hard to see. Which means the fight has moved. It’s not really about who’s smartest any more. It’s about price, memory, and who lets you take the model home. Google’s betting you’ll pay rent once you’ve settled in. Meta’s betting you’d rather own the place. And we’ll find out who read that right around the first of December, when that price doubles. Anyway, that’s it for this one. Tell me in the comments which one you’re actually using. Full comparisons and pricing are on AI Tools Review dot co dot UK. See you in the next one.